2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. All advertising materials mentioning features or use of this software
14 * must display the following acknowledgement:
15 * This product includes software developed by the University of
16 * California, Berkeley and its contributors.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
34 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
35 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.33 2004/06/07 02:36:22 dillon Exp $
38 #include "opt_compat.h"
39 #include "opt_inet6.h"
40 #include "opt_ipsec.h"
41 #include "opt_tcpdebug.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/callout.h>
46 #include <sys/kernel.h>
47 #include <sys/sysctl.h>
48 #include <sys/malloc.h>
49 #include <sys/mpipe.h>
52 #include <sys/domain.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/protosw.h>
58 #include <sys/random.h>
59 #include <sys/in_cksum.h>
61 #include <vm/vm_zone.h>
63 #include <net/route.h>
65 #include <net/netisr.h>
68 #include <netinet/in.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/in_pcb.h>
73 #include <netinet6/in6_pcb.h>
74 #include <netinet/in_var.h>
75 #include <netinet/ip_var.h>
76 #include <netinet6/ip6_var.h>
77 #include <netinet/tcp.h>
78 #include <netinet/tcp_fsm.h>
79 #include <netinet/tcp_seq.h>
80 #include <netinet/tcp_timer.h>
81 #include <netinet/tcp_var.h>
82 #include <netinet6/tcp6_var.h>
83 #include <netinet/tcpip.h>
85 #include <netinet/tcp_debug.h>
87 #include <netinet6/ip6protosw.h>
90 #include <netinet6/ipsec.h>
92 #include <netinet6/ipsec6.h>
97 #include <netipsec/ipsec.h>
99 #include <netipsec/ipsec6.h>
106 #include <sys/msgport2.h>
108 #include <machine/smp.h>
110 int tcp_mssdflt = TCP_MSS;
111 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
112 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
115 int tcp_v6mssdflt = TCP6_MSS;
116 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
117 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
121 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
122 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
123 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
126 int tcp_do_rfc1323 = 1;
127 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
128 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
130 int tcp_do_rfc1644 = 0;
131 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
132 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
134 static int tcp_tcbhashsize = 0;
135 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
136 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
138 static int do_tcpdrain = 1;
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
140 "Enable tcp_drain routine for extra help when low on mbufs");
143 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
144 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
146 static int icmp_may_rst = 1;
147 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
148 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
150 static int tcp_isn_reseed_interval = 0;
151 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
152 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
155 * TCP bandwidth limiting sysctls. Note that the default lower bound of
156 * 1024 exists only for debugging. A good production default would be
157 * something like 6100.
159 static int tcp_inflight_enable = 0;
160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
161 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
163 static int tcp_inflight_debug = 0;
164 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
165 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
167 static int tcp_inflight_min = 6144;
168 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
169 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
171 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
172 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
173 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
175 static int tcp_inflight_stab = 20;
176 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
177 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
179 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
180 static struct malloc_pipe tcptemp_mpipe;
182 static void tcp_cleartaocache (void);
183 static void tcp_notify (struct inpcb *, int);
185 struct tcp_stats tcpstats_ary[MAXCPU];
188 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
192 for (cpu = 0; cpu < ncpus; ++cpu) {
193 if ((error = SYSCTL_OUT(req, (void *)&tcpstats_ary[cpu],
194 sizeof(struct tcp_stats))))
196 if ((error = SYSCTL_IN(req, (void *)&tcpstats_ary[cpu],
197 sizeof(struct tcp_stats))))
203 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
204 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
206 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
207 &tcpstat, tcp_stats, "TCP statistics");
211 * Target size of TCP PCB hash tables. Must be a power of two.
213 * Note that this can be overridden by the kernel environment
214 * variable net.inet.tcp.tcbhashsize
217 #define TCBHASHSIZE 512
221 * This is the actual shape of what we allocate using the zone
222 * allocator. Doing it this way allows us to protect both structures
223 * using the same generation count, and also eliminates the overhead
224 * of allocating tcpcbs separately. By hiding the structure here,
225 * we avoid changing most of the rest of the code (although it needs
226 * to be changed, eventually, for greater efficiency).
229 #define ALIGNM1 (ALIGNMENT - 1)
233 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
236 struct callout inp_tp_rexmt, inp_tp_persist, inp_tp_keep, inp_tp_2msl;
237 struct callout inp_tp_delack;
248 struct inpcbporthead *porthashbase;
250 struct vm_zone *ipi_zone;
251 int hashsize = TCBHASHSIZE;
255 * note: tcptemp is used for keepalives, and it is ok for an
256 * allocation to fail so do not specify MPF_INT.
258 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
264 tcp_delacktime = TCPTV_DELACK;
265 tcp_keepinit = TCPTV_KEEP_INIT;
266 tcp_keepidle = TCPTV_KEEP_IDLE;
267 tcp_keepintvl = TCPTV_KEEPINTVL;
268 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
270 tcp_rexmit_min = TCPTV_MIN;
271 tcp_rexmit_slop = TCPTV_CPU_VAR;
273 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
274 if (!powerof2(hashsize)) {
275 printf("WARNING: TCB hash size not a power of 2\n");
276 hashsize = 512; /* safe default */
278 tcp_tcbhashsize = hashsize;
279 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
280 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
283 for (cpu = 0; cpu < ncpus2; cpu++) {
284 in_pcbinfo_init(&tcbinfo[cpu]);
285 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
286 &tcbinfo[cpu].hashmask);
287 tcbinfo[cpu].porthashbase = porthashbase;
288 tcbinfo[cpu].porthashmask = porthashmask;
289 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
290 &tcbinfo[cpu].wildcardhashmask);
291 tcbinfo[cpu].ipi_zone = ipi_zone;
294 tcp_reass_maxseg = nmbclusters / 16;
295 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
298 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
300 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
302 if (max_protohdr < TCP_MINPROTOHDR)
303 max_protohdr = TCP_MINPROTOHDR;
304 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
306 #undef TCP_MINPROTOHDR
309 * Initialize TCP statistics.
311 * It is layed out as an array which is has one element for UP,
312 * and SMP_MAXCPU elements for SMP. This allows us to retain
313 * the access mechanism from userland for both UP and SMP.
316 for (cpu = 0; cpu < ncpus; ++cpu) {
317 bzero(&tcpstats_ary[cpu], sizeof(struct tcp_stats));
320 bzero(&tcpstat, sizeof(struct tcp_stats));
328 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
329 * tcp_template used to store this data in mbufs, but we now recopy it out
330 * of the tcpcb each time to conserve mbufs.
333 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
335 struct inpcb *inp = tp->t_inpcb;
336 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
339 if (inp->inp_vflag & INP_IPV6) {
342 ip6 = (struct ip6_hdr *)ip_ptr;
343 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
344 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
345 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
346 (IPV6_VERSION & IPV6_VERSION_MASK);
347 ip6->ip6_nxt = IPPROTO_TCP;
348 ip6->ip6_plen = sizeof(struct tcphdr);
349 ip6->ip6_src = inp->in6p_laddr;
350 ip6->ip6_dst = inp->in6p_faddr;
355 struct ip *ip = (struct ip *) ip_ptr;
357 ip->ip_vhl = IP_VHL_BORING;
364 ip->ip_p = IPPROTO_TCP;
365 ip->ip_src = inp->inp_laddr;
366 ip->ip_dst = inp->inp_faddr;
367 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
369 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
372 tcp_hdr->th_sport = inp->inp_lport;
373 tcp_hdr->th_dport = inp->inp_fport;
378 tcp_hdr->th_flags = 0;
384 * Create template to be used to send tcp packets on a connection.
385 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
386 * use for this function is in keepalives, which use tcp_respond.
389 tcp_maketemplate(struct tcpcb *tp)
393 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
395 tcp_fillheaders(tp, (void *)&tmp->tt_ipgen, (void *)&tmp->tt_t);
400 tcp_freetemplate(struct tcptemp *tmp)
402 mpipe_free(&tcptemp_mpipe, tmp);
406 * Send a single message to the TCP at address specified by
407 * the given TCP/IP header. If m == NULL, then we make a copy
408 * of the tcpiphdr at ti and send directly to the addressed host.
409 * This is used to force keep alive messages out using the TCP
410 * template for a connection. If flags are given then we send
411 * a message back to the TCP which originated the * segment ti,
412 * and discard the mbuf containing it and any other attached mbufs.
414 * In any case the ack and sequence number of the transmitted
415 * segment are as specified by the parameters.
417 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
420 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
421 tcp_seq ack, tcp_seq seq, int flags)
425 struct route *ro = NULL;
427 struct ip *ip = ipgen;
430 struct route_in6 *ro6 = NULL;
431 struct route_in6 sro6;
432 struct ip6_hdr *ip6 = ipgen;
434 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
436 const boolean_t isipv6 = FALSE;
440 if (!(flags & TH_RST)) {
441 win = sbspace(&tp->t_inpcb->inp_socket->so_rcv);
442 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
443 win = (long)TCP_MAXWIN << tp->rcv_scale;
446 ro6 = &tp->t_inpcb->in6p_route;
448 ro = &tp->t_inpcb->inp_route;
452 bzero(ro6, sizeof *ro6);
455 bzero(ro, sizeof *ro);
459 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
463 m->m_data += max_linkhdr;
465 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
466 ip6 = mtod(m, struct ip6_hdr *);
467 nth = (struct tcphdr *)(ip6 + 1);
469 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
470 ip = mtod(m, struct ip *);
471 nth = (struct tcphdr *)(ip + 1);
473 bcopy(th, nth, sizeof(struct tcphdr));
478 m->m_data = (caddr_t)ipgen;
479 /* m_len is set later */
481 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
483 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
484 nth = (struct tcphdr *)(ip6 + 1);
486 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
487 nth = (struct tcphdr *)(ip + 1);
491 * this is usually a case when an extension header
492 * exists between the IPv6 header and the
495 nth->th_sport = th->th_sport;
496 nth->th_dport = th->th_dport;
498 xchg(nth->th_dport, nth->th_sport, n_short);
503 ip6->ip6_vfc = IPV6_VERSION;
504 ip6->ip6_nxt = IPPROTO_TCP;
505 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
506 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
508 tlen += sizeof(struct tcpiphdr);
510 ip->ip_ttl = ip_defttl;
513 m->m_pkthdr.len = tlen;
514 m->m_pkthdr.rcvif = (struct ifnet *) NULL;
515 nth->th_seq = htonl(seq);
516 nth->th_ack = htonl(ack);
518 nth->th_off = sizeof(struct tcphdr) >> 2;
519 nth->th_flags = flags;
521 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
523 nth->th_win = htons((u_short)win);
527 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
528 sizeof(struct ip6_hdr),
529 tlen - sizeof(struct ip6_hdr));
530 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
531 (ro6 && ro6->ro_rt) ?
532 ro6->ro_rt->rt_ifp : NULL);
534 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
535 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
536 m->m_pkthdr.csum_flags = CSUM_TCP;
537 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
540 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
541 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
544 (void)ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
545 tp ? tp->t_inpcb : NULL);
546 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
551 (void)ip_output(m, NULL, ro, ipflags, NULL,
552 tp ? tp->t_inpcb : NULL);
553 if ((ro == &sro) && (ro->ro_rt != NULL)) {
561 * Create a new TCP control block, making an
562 * empty reassembly queue and hooking it to the argument
563 * protocol control block. The `inp' parameter must have
564 * come from the zone allocator set up in tcp_init().
567 tcp_newtcpcb(struct inpcb *inp)
572 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
574 const boolean_t isipv6 = FALSE;
577 it = (struct inp_tp *)inp;
579 bzero(tp, sizeof(struct tcpcb));
580 LIST_INIT(&tp->t_segq);
581 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
583 /* Set up our timeouts. */
584 callout_init(tp->tt_rexmt = &it->inp_tp_rexmt);
585 callout_init(tp->tt_persist = &it->inp_tp_persist);
586 callout_init(tp->tt_keep = &it->inp_tp_keep);
587 callout_init(tp->tt_2msl = &it->inp_tp_2msl);
588 callout_init(tp->tt_delack = &it->inp_tp_delack);
591 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
593 tp->t_flags |= TF_REQ_CC;
594 tp->t_inpcb = inp; /* XXX */
596 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
597 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
598 * reasonable initial retransmit time.
600 tp->t_srtt = TCPTV_SRTTBASE;
602 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
603 tp->t_rttmin = tcp_rexmit_min;
604 tp->t_rxtcur = TCPTV_RTOBASE;
605 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
606 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
607 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
608 tp->t_rcvtime = ticks;
609 tp->t_bw_rtttime = ticks;
611 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
612 * because the socket may be bound to an IPv6 wildcard address,
613 * which may match an IPv4-mapped IPv6 address.
615 inp->inp_ip_ttl = ip_defttl;
616 inp->inp_ppcb = (caddr_t)tp;
617 return (tp); /* XXX */
621 * Drop a TCP connection, reporting the specified error.
622 * If connection is synchronized, then send a RST to peer.
625 tcp_drop(struct tcpcb *tp, int errno)
627 struct socket *so = tp->t_inpcb->inp_socket;
629 if (TCPS_HAVERCVDSYN(tp->t_state)) {
630 tp->t_state = TCPS_CLOSED;
631 (void) tcp_output(tp);
632 tcpstat.tcps_drops++;
634 tcpstat.tcps_conndrops++;
635 if (errno == ETIMEDOUT && tp->t_softerror)
636 errno = tp->t_softerror;
637 so->so_error = errno;
638 return (tcp_close(tp));
642 struct netmsg_remwildcard {
643 struct lwkt_msg nm_lmsg;
644 struct inpcb *nm_inp;
645 struct inpcbinfo *nm_pcbinfo;
649 in_pcbremwildcardhash_handler(struct lwkt_msg *msg0)
651 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
653 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
654 lwkt_replymsg(&msg->nm_lmsg, 0);
660 * Close a TCP control block:
661 * discard all space held by the tcp
662 * discard internet protocol block
663 * wake up any sleepers
666 tcp_close(struct tcpcb *tp)
669 struct inpcb *inp = tp->t_inpcb;
670 struct socket *so = inp->inp_socket;
672 boolean_t dosavessthresh;
677 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
679 const boolean_t isipv6 = FALSE;
683 * Make sure that all of our timers are stopped before we
686 callout_stop(tp->tt_rexmt);
687 callout_stop(tp->tt_persist);
688 callout_stop(tp->tt_keep);
689 callout_stop(tp->tt_2msl);
690 callout_stop(tp->tt_delack);
693 * If we got enough samples through the srtt filter,
694 * save the rtt and rttvar in the routing entry.
695 * 'Enough' is arbitrarily defined as the 16 samples.
696 * 16 samples is enough for the srtt filter to converge
697 * to within 5% of the correct value; fewer samples and
698 * we could save a very bogus rtt.
700 * Don't update the default route's characteristics and don't
701 * update anything that the user "locked".
703 if (tp->t_rttupdated >= 16) {
707 struct sockaddr_in6 *sin6;
709 if ((rt = inp->in6p_route.ro_rt) == NULL)
711 sin6 = (struct sockaddr_in6 *)rt_key(rt);
712 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
715 if ((rt = inp->inp_route.ro_rt) == NULL ||
716 ((struct sockaddr_in *)rt_key(rt))->
717 sin_addr.s_addr == INADDR_ANY)
720 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
721 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
722 if (rt->rt_rmx.rmx_rtt && i)
724 * filter this update to half the old & half
725 * the new values, converting scale.
726 * See route.h and tcp_var.h for a
727 * description of the scaling constants.
730 (rt->rt_rmx.rmx_rtt + i) / 2;
732 rt->rt_rmx.rmx_rtt = i;
733 tcpstat.tcps_cachedrtt++;
735 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
737 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
738 if (rt->rt_rmx.rmx_rttvar && i)
739 rt->rt_rmx.rmx_rttvar =
740 (rt->rt_rmx.rmx_rttvar + i) / 2;
742 rt->rt_rmx.rmx_rttvar = i;
743 tcpstat.tcps_cachedrttvar++;
746 * The old comment here said:
747 * update the pipelimit (ssthresh) if it has been updated
748 * already or if a pipesize was specified & the threshhold
749 * got below half the pipesize. I.e., wait for bad news
750 * before we start updating, then update on both good
753 * But we want to save the ssthresh even if no pipesize is
754 * specified explicitly in the route, because such
755 * connections still have an implicit pipesize specified
756 * by the global tcp_sendspace. In the absence of a reliable
757 * way to calculate the pipesize, it will have to do.
759 i = tp->snd_ssthresh;
760 if (rt->rt_rmx.rmx_sendpipe != 0)
761 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
763 dosavessthresh = (i < so->so_snd.sb_hiwat/2);
764 if (dosavessthresh ||
765 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
766 (rt->rt_rmx.rmx_ssthresh != 0))) {
768 * convert the limit from user data bytes to
769 * packets then to packet data bytes.
771 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
776 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
777 sizeof(struct tcpiphdr));
778 if (rt->rt_rmx.rmx_ssthresh)
779 rt->rt_rmx.rmx_ssthresh =
780 (rt->rt_rmx.rmx_ssthresh + i) / 2;
782 rt->rt_rmx.rmx_ssthresh = i;
783 tcpstat.tcps_cachedssthresh++;
788 /* free the reassembly queue, if any */
789 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
790 LIST_REMOVE(q, tqe_q);
795 inp->inp_ppcb = NULL;
796 soisdisconnected(so);
799 if (inp->inp_flags & INP_WILDCARD_MP) {
800 for (cpu = 0; cpu < ncpus2; cpu ++) {
801 struct netmsg_remwildcard *msg;
803 msg = malloc(sizeof(struct netmsg_remwildcard),
804 M_LWKTMSG, M_INTWAIT);
805 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
806 lwkt_cmd_func(in_pcbremwildcardhash_handler),
809 msg->nm_pcbinfo = &tcbinfo[cpu];
810 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
816 if (INP_CHECK_SOCKAF(so, AF_INET6))
821 tcpstat.tcps_closed++;
826 tcp_drain_oncpu(struct inpcbhead *head)
830 struct tseg_qent *te;
832 LIST_FOREACH(inpb, head, inp_list) {
833 if (inpb->inp_flags & INP_PLACEMARKER)
835 if ((tcpb = intotcpcb(inpb))) {
836 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
837 LIST_REMOVE(te, tqe_q);
847 struct netmsg_tcp_drain {
848 struct lwkt_msg nm_lmsg;
849 struct inpcbhead *nm_head;
853 tcp_drain_handler(lwkt_msg_t lmsg)
855 struct netmsg_tcp_drain *nm = (void *)lmsg;
857 tcp_drain_oncpu(nm->nm_head);
858 lwkt_replymsg(lmsg, 0);
874 * Walk the tcpbs, if existing, and flush the reassembly queue,
876 * XXX: The "Net/3" implementation doesn't imply that the TCP
877 * reassembly queue should be flushed, but in a situation
878 * where we're really low on mbufs, this is potentially
882 for (cpu = 0; cpu < ncpus2; cpu++) {
883 struct netmsg_tcp_drain *msg;
885 if (cpu == mycpu->gd_cpuid) {
886 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
888 msg = malloc(sizeof(struct netmsg_tcp_drain),
889 M_LWKTMSG, M_NOWAIT);
892 lwkt_initmsg(&msg->nm_lmsg, &netisr_afree_rport, 0,
893 lwkt_cmd_func(tcp_drain_handler),
895 msg->nm_head = &tcbinfo[cpu].pcblisthead;
896 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_lmsg);
900 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
905 * Notify a tcp user of an asynchronous error;
906 * store error as soft error, but wake up user
907 * (for now, won't do anything until can select for soft error).
909 * Do not wake up user since there currently is no mechanism for
910 * reporting soft errors (yet - a kqueue filter may be added).
913 tcp_notify(struct inpcb *inp, int error)
915 struct tcpcb *tp = intotcpcb(inp);
918 * Ignore some errors if we are hooked up.
919 * If connection hasn't completed, has retransmitted several times,
920 * and receives a second error, give up now. This is better
921 * than waiting a long time to establish a connection that
922 * can never complete.
924 if (tp->t_state == TCPS_ESTABLISHED &&
925 (error == EHOSTUNREACH || error == ENETUNREACH ||
926 error == EHOSTDOWN)) {
928 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
932 tp->t_softerror = error;
934 wakeup((caddr_t) &so->so_timeo);
941 tcp_pcblist(SYSCTL_HANDLER_ARGS)
944 struct inpcb *marker;
955 * The process of preparing the TCB list is too time-consuming and
956 * resource-intensive to repeat twice on every request.
958 if (req->oldptr == NULL) {
959 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
960 gd = globaldata_find(ccpu);
961 n += tcbinfo[gd->gd_cpuid].ipi_count;
963 req->oldidx = 2 * ncpus * (sizeof xig) +
964 (n + n/8) * sizeof(struct xtcpcb);
968 if (req->newptr != NULL)
971 marker = malloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
972 marker->inp_flags |= INP_PLACEMARKER;
975 * OK, now we're committed to doing something. Run the inpcb list
976 * for each cpu in the system and construct the output. Use a
977 * list placemarker to deal with list changes occuring during
978 * copyout blockages (but otherwise depend on being on the correct
979 * cpu to avoid races).
981 origcpu = mycpu->gd_cpuid;
982 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
988 cpu_id = (origcpu + ccpu) % ncpus;
989 if ((smp_active_mask & (1 << cpu_id)) == 0)
991 rgd = globaldata_find(cpu_id);
992 lwkt_setcpu_self(rgd);
994 /* indicate change of CPU */
997 gencnt = tcbinfo[cpu_id].ipi_gencnt;
998 n = tcbinfo[cpu_id].ipi_count;
1000 xig.xig_len = sizeof xig;
1002 xig.xig_gen = gencnt;
1003 xig.xig_sogen = so_gencnt;
1004 xig.xig_cpu = cpu_id;
1005 error = SYSCTL_OUT(req, &xig, sizeof xig);
1009 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1011 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1013 * process a snapshot of pcbs, ignoring placemarkers
1014 * and using our own to allow SYSCTL_OUT to block.
1016 LIST_REMOVE(marker, inp_list);
1017 LIST_INSERT_AFTER(inp, marker, inp_list);
1019 if (inp->inp_flags & INP_PLACEMARKER)
1021 if (inp->inp_gencnt > gencnt)
1023 if (prison_xinpcb(req->td, inp))
1026 xt.xt_len = sizeof xt;
1027 bcopy(inp, &xt.xt_inp, sizeof *inp);
1028 inp_ppcb = inp->inp_ppcb;
1029 if (inp_ppcb != NULL)
1030 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1032 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1033 if (inp->inp_socket)
1034 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1035 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1039 LIST_REMOVE(marker, inp_list);
1040 if (error == 0 && i < n) {
1041 bzero(&xt, sizeof(xt));
1042 xt.xt_len = sizeof(xt);
1044 error = SYSCTL_OUT(req, &xt, sizeof (xt));
1052 * Give the user an updated idea of our state.
1053 * If the generation differs from what we told
1054 * her before, she knows that something happened
1055 * while we were processing this request, and it
1056 * might be necessary to retry.
1058 xig.xig_gen = tcbinfo[cpu_id].ipi_gencnt;
1059 xig.xig_sogen = so_gencnt;
1060 xig.xig_count = tcbinfo[cpu_id].ipi_count;
1061 error = SYSCTL_OUT(req, &xig, sizeof xig);
1066 * Make sure we are on the same cpu we were on originally, since
1067 * higher level callers expect this. Also don't pollute caches with
1068 * migrated userland data by (eventually) returning to userland
1069 * on a different cpu.
1071 lwkt_setcpu_self(globaldata_find(origcpu));
1072 free(marker, M_TEMP);
1076 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1077 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1080 tcp_getcred(SYSCTL_HANDLER_ARGS)
1082 struct sockaddr_in addrs[2];
1087 error = suser(req->td);
1090 error = SYSCTL_IN(req, addrs, sizeof addrs);
1095 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1096 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1097 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1098 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1099 if (inp == NULL || inp->inp_socket == NULL) {
1103 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1109 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1110 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1114 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1116 struct sockaddr_in6 addrs[2];
1119 boolean_t mapped = FALSE;
1121 error = suser(req->td);
1124 error = SYSCTL_IN(req, addrs, sizeof addrs);
1127 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1128 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1135 inp = in_pcblookup_hash(&tcbinfo[0],
1136 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1138 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1142 inp = in6_pcblookup_hash(&tcbinfo[0],
1143 &addrs[1].sin6_addr, addrs[1].sin6_port,
1144 &addrs[0].sin6_addr, addrs[0].sin6_port,
1147 if (inp == NULL || inp->inp_socket == NULL) {
1151 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1157 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1159 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1163 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1165 struct ip *ip = vip;
1167 struct in_addr faddr;
1170 void (*notify)(struct inpcb *, int) = tcp_notify;
1175 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1176 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1179 if (cmd == PRC_QUENCH)
1180 notify = tcp_quench;
1181 else if (icmp_may_rst &&
1182 (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT ||
1183 cmd == PRC_TIMXCEED_INTRANS) &&
1185 notify = tcp_drop_syn_sent;
1186 else if (cmd == PRC_MSGSIZE)
1187 notify = tcp_mtudisc;
1188 else if (PRC_IS_REDIRECT(cmd)) {
1190 notify = in_rtchange;
1191 } else if (cmd == PRC_HOSTDEAD)
1193 else if ((unsigned)cmd > PRC_NCMDS || inetctlerrmap[cmd] == 0)
1197 th = (struct tcphdr *)((caddr_t)ip +
1198 (IP_VHL_HL(ip->ip_vhl) << 2));
1199 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1200 ip->ip_src.s_addr, th->th_sport);
1201 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1202 ip->ip_src, th->th_sport, 0, NULL);
1203 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1204 icmp_seq = htonl(th->th_seq);
1205 tp = intotcpcb(inp);
1206 if (SEQ_GEQ(icmp_seq, tp->snd_una) &&
1207 SEQ_LT(icmp_seq, tp->snd_max))
1208 (*notify)(inp, inetctlerrmap[cmd]);
1210 struct in_conninfo inc;
1212 inc.inc_fport = th->th_dport;
1213 inc.inc_lport = th->th_sport;
1214 inc.inc_faddr = faddr;
1215 inc.inc_laddr = ip->ip_src;
1219 syncache_unreach(&inc, th);
1223 for (cpu = 0; cpu < ncpus2; cpu++) {
1224 in_pcbnotifyall(&tcbinfo[cpu].pcblisthead, faddr,
1225 inetctlerrmap[cmd], notify);
1232 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1235 void (*notify) (struct inpcb *, int) = tcp_notify;
1236 struct ip6_hdr *ip6;
1238 struct ip6ctlparam *ip6cp = NULL;
1239 const struct sockaddr_in6 *sa6_src = NULL;
1241 struct tcp_portonly {
1246 if (sa->sa_family != AF_INET6 ||
1247 sa->sa_len != sizeof(struct sockaddr_in6))
1250 if (cmd == PRC_QUENCH)
1251 notify = tcp_quench;
1252 else if (cmd == PRC_MSGSIZE)
1253 notify = tcp_mtudisc;
1254 else if (!PRC_IS_REDIRECT(cmd) &&
1255 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0))
1258 /* if the parameter is from icmp6, decode it. */
1260 ip6cp = (struct ip6ctlparam *)d;
1262 ip6 = ip6cp->ip6c_ip6;
1263 off = ip6cp->ip6c_off;
1264 sa6_src = ip6cp->ip6c_src;
1268 off = 0; /* fool gcc */
1273 struct in_conninfo inc;
1275 * XXX: We assume that when IPV6 is non NULL,
1276 * M and OFF are valid.
1279 /* check if we can safely examine src and dst ports */
1280 if (m->m_pkthdr.len < off + sizeof *thp)
1283 bzero(&th, sizeof th);
1284 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1286 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1287 (struct sockaddr *)ip6cp->ip6c_src,
1288 th.th_sport, cmd, notify);
1290 inc.inc_fport = th.th_dport;
1291 inc.inc_lport = th.th_sport;
1292 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1293 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1295 syncache_unreach(&inc, &th);
1297 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1298 (const struct sockaddr *)sa6_src, 0, cmd, notify);
1303 * Following is where TCP initial sequence number generation occurs.
1305 * There are two places where we must use initial sequence numbers:
1306 * 1. In SYN-ACK packets.
1307 * 2. In SYN packets.
1309 * All ISNs for SYN-ACK packets are generated by the syncache. See
1310 * tcp_syncache.c for details.
1312 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1313 * depends on this property. In addition, these ISNs should be
1314 * unguessable so as to prevent connection hijacking. To satisfy
1315 * the requirements of this situation, the algorithm outlined in
1316 * RFC 1948 is used to generate sequence numbers.
1318 * Implementation details:
1320 * Time is based off the system timer, and is corrected so that it
1321 * increases by one megabyte per second. This allows for proper
1322 * recycling on high speed LANs while still leaving over an hour
1325 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1326 * between seeding of isn_secret. This is normally set to zero,
1327 * as reseeding should not be necessary.
1331 #define ISN_BYTES_PER_SECOND 1048576
1333 u_char isn_secret[32];
1334 int isn_last_reseed;
1338 tcp_new_isn(struct tcpcb *tp)
1340 u_int32_t md5_buffer[4];
1343 /* Seed if this is the first use, reseed if requested. */
1344 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1345 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1347 read_random_unlimited(&isn_secret, sizeof isn_secret);
1348 isn_last_reseed = ticks;
1351 /* Compute the md5 hash and return the ISN. */
1353 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1354 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1356 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1357 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1358 sizeof(struct in6_addr));
1359 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1360 sizeof(struct in6_addr));
1364 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1365 sizeof(struct in_addr));
1366 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1367 sizeof(struct in_addr));
1369 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1370 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1371 new_isn = (tcp_seq) md5_buffer[0];
1372 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1377 * When a source quench is received, close congestion window
1378 * to one segment. We will gradually open it again as we proceed.
1381 tcp_quench(struct inpcb *inp, int errno)
1383 struct tcpcb *tp = intotcpcb(inp);
1386 tp->snd_cwnd = tp->t_maxseg;
1390 * When a specific ICMP unreachable message is received and the
1391 * connection state is SYN-SENT, drop the connection. This behavior
1392 * is controlled by the icmp_may_rst sysctl.
1395 tcp_drop_syn_sent(struct inpcb *inp, int errno)
1397 struct tcpcb *tp = intotcpcb(inp);
1399 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1400 tcp_drop(tp, errno);
1404 * When `need fragmentation' ICMP is received, update our idea of the MSS
1405 * based on the new value in the route. Also nudge TCP to send something,
1406 * since we know the packet we just sent was dropped.
1407 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1410 tcp_mtudisc(struct inpcb *inp, int errno)
1412 struct tcpcb *tp = intotcpcb(inp);
1414 struct rmxp_tao *taop;
1415 struct socket *so = inp->inp_socket;
1419 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1421 const boolean_t isipv6 = FALSE;
1426 rt = tcp_rtlookup6(&inp->inp_inc);
1428 rt = tcp_rtlookup(&inp->inp_inc);
1429 if (rt == NULL || rt->rt_rmx.rmx_mtu == 0) {
1430 tp->t_maxopd = tp->t_maxseg =
1431 isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
1434 taop = rmx_taop(rt->rt_rmx);
1435 offered = taop->tao_mssopt;
1436 mss = rt->rt_rmx.rmx_mtu -
1438 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1439 sizeof(struct tcpiphdr));
1442 mss = min(mss, offered);
1444 * XXX - The above conditional probably violates the TCP
1445 * spec. The problem is that, since we don't know the
1446 * other end's MSS, we are supposed to use a conservative
1447 * default. But, if we do that, then MTU discovery will
1448 * never actually take place, because the conservative
1449 * default is much less than the MTUs typically seen
1450 * on the Internet today. For the moment, we'll sweep
1451 * this under the carpet.
1453 * The conservative default might not actually be a problem
1454 * if the only case this occurs is when sending an initial
1455 * SYN with options and data to a host we've never talked
1456 * to before. Then, they will reply with an MSS value which
1457 * will get recorded and the new parameters should get
1458 * recomputed. For Further Study.
1460 if (tp->t_maxopd <= mss)
1464 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1465 (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)
1466 mss -= TCPOLEN_TSTAMP_APPA;
1467 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
1468 (tp->t_flags & TF_RCVD_CC) == TF_RCVD_CC)
1469 mss -= TCPOLEN_CC_APPA;
1470 #if (MCLBYTES & (MCLBYTES - 1)) == 0
1472 mss &= ~(MCLBYTES - 1);
1475 mss = mss / MCLBYTES * MCLBYTES;
1477 if (so->so_snd.sb_hiwat < mss)
1478 mss = so->so_snd.sb_hiwat;
1482 tcpstat.tcps_mturesent++;
1484 tp->snd_nxt = tp->snd_una;
1490 * Look-up the routing entry to the peer of this inpcb. If no route
1491 * is found and it cannot be allocated the return NULL. This routine
1492 * is called by TCP routines that access the rmx structure and by tcp_mss
1493 * to get the interface MTU.
1496 tcp_rtlookup(struct in_conninfo *inc)
1501 ro = &inc->inc_route;
1503 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1504 /* No route yet, so try to acquire one */
1505 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1506 ro->ro_dst.sa_family = AF_INET;
1507 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1508 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1519 tcp_rtlookup6(struct in_conninfo *inc)
1521 struct route_in6 *ro6;
1524 ro6 = &inc->inc6_route;
1526 if (rt == NULL || !(rt->rt_flags & RTF_UP)) {
1527 /* No route yet, so try to acquire one */
1528 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1529 ro6->ro_dst.sin6_family = AF_INET6;
1530 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1531 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1532 rtalloc((struct route *)ro6);
1541 /* compute ESP/AH header size for TCP, including outer IP header. */
1543 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1551 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1553 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1558 if (inp->inp_vflag & INP_IPV6) {
1559 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1561 th = (struct tcphdr *)(ip6 + 1);
1562 m->m_pkthdr.len = m->m_len =
1563 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1564 tcp_fillheaders(tp, ip6, th);
1565 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1569 ip = mtod(m, struct ip *);
1570 th = (struct tcphdr *)(ip + 1);
1571 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1572 tcp_fillheaders(tp, ip, th);
1573 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1582 * Return a pointer to the cached information about the remote host.
1583 * The cached information is stored in the protocol specific part of
1584 * the route metrics.
1587 tcp_gettaocache(struct in_conninfo *inc)
1592 if (inc->inc_isipv6)
1593 rt = tcp_rtlookup6(inc);
1596 rt = tcp_rtlookup(inc);
1598 /* Make sure this is a host route and is up. */
1600 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1603 return (rmx_taop(rt->rt_rmx));
1607 * Clear all the TAO cache entries, called from tcp_init.
1610 * This routine is just an empty one, because we assume that the routing
1611 * routing tables are initialized at the same time when TCP, so there is
1612 * nothing in the cache left over.
1620 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1622 * This code attempts to calculate the bandwidth-delay product as a
1623 * means of determining the optimal window size to maximize bandwidth,
1624 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1625 * routers. This code also does a fairly good job keeping RTTs in check
1626 * across slow links like modems. We implement an algorithm which is very
1627 * similar (but not meant to be) TCP/Vegas. The code operates on the
1628 * transmitter side of a TCP connection and so only effects the transmit
1629 * side of the connection.
1631 * BACKGROUND: TCP makes no provision for the management of buffer space
1632 * at the end points or at the intermediate routers and switches. A TCP
1633 * stream, whether using NewReno or not, will eventually buffer as
1634 * many packets as it is able and the only reason this typically works is
1635 * due to the fairly small default buffers made available for a connection
1636 * (typicaly 16K or 32K). As machines use larger windows and/or window
1637 * scaling it is now fairly easy for even a single TCP connection to blow-out
1638 * all available buffer space not only on the local interface, but on
1639 * intermediate routers and switches as well. NewReno makes a misguided
1640 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1641 * then backing off, then steadily increasing the window again until another
1642 * failure occurs, ad-infinitum. This results in terrible oscillation that
1643 * is only made worse as network loads increase and the idea of intentionally
1644 * blowing out network buffers is, frankly, a terrible way to manage network
1647 * It is far better to limit the transmit window prior to the failure
1648 * condition being achieved. There are two general ways to do this: First
1649 * you can 'scan' through different transmit window sizes and locate the
1650 * point where the RTT stops increasing, indicating that you have filled the
1651 * pipe, then scan backwards until you note that RTT stops decreasing, then
1652 * repeat ad-infinitum. This method works in principle but has severe
1653 * implementation issues due to RTT variances, timer granularity, and
1654 * instability in the algorithm which can lead to many false positives and
1655 * create oscillations as well as interact badly with other TCP streams
1656 * implementing the same algorithm.
1658 * The second method is to limit the window to the bandwidth delay product
1659 * of the link. This is the method we implement. RTT variances and our
1660 * own manipulation of the congestion window, bwnd, can potentially
1661 * destabilize the algorithm. For this reason we have to stabilize the
1662 * elements used to calculate the window. We do this by using the minimum
1663 * observed RTT, the long term average of the observed bandwidth, and
1664 * by adding two segments worth of slop. It isn't perfect but it is able
1665 * to react to changing conditions and gives us a very stable basis on
1666 * which to extend the algorithm.
1669 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1676 * If inflight_enable is disabled in the middle of a tcp connection,
1677 * make sure snd_bwnd is effectively disabled.
1679 if (!tcp_inflight_enable) {
1680 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1681 tp->snd_bandwidth = 0;
1686 * Figure out the bandwidth. Due to the tick granularity this
1687 * is a very rough number and it MUST be averaged over a fairly
1688 * long period of time. XXX we need to take into account a link
1689 * that is not using all available bandwidth, but for now our
1690 * slop will ramp us up if this case occurs and the bandwidth later
1693 * Note: if ticks rollover 'bw' may wind up negative. We must
1694 * effectively reset t_bw_rtttime for this case.
1697 if ((u_int)(save_ticks - tp->t_bw_rtttime) < 1)
1700 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz /
1701 (save_ticks - tp->t_bw_rtttime);
1702 tp->t_bw_rtttime = save_ticks;
1703 tp->t_bw_rtseq = ack_seq;
1704 if (tp->t_bw_rtttime == 0 || (int)bw < 0)
1706 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1708 tp->snd_bandwidth = bw;
1711 * Calculate the semi-static bandwidth delay product, plus two maximal
1712 * segments. The additional slop puts us squarely in the sweet
1713 * spot and also handles the bandwidth run-up case. Without the
1714 * slop we could be locking ourselves into a lower bandwidth.
1716 * Situations Handled:
1717 * (1) Prevents over-queueing of packets on LANs, especially on
1718 * high speed LANs, allowing larger TCP buffers to be
1719 * specified, and also does a good job preventing
1720 * over-queueing of packets over choke points like modems
1721 * (at least for the transmit side).
1723 * (2) Is able to handle changing network loads (bandwidth
1724 * drops so bwnd drops, bandwidth increases so bwnd
1727 * (3) Theoretically should stabilize in the face of multiple
1728 * connections implementing the same algorithm (this may need
1731 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1732 * be adjusted with a sysctl but typically only needs to be on
1733 * very slow connections. A value no smaller then 5 should
1734 * be used, but only reduce this default if you have no other
1738 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1739 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1740 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1743 if (tcp_inflight_debug > 0) {
1745 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1747 printf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1748 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1751 if ((long)bwnd < tcp_inflight_min)
1752 bwnd = tcp_inflight_min;
1753 if (bwnd > tcp_inflight_max)
1754 bwnd = tcp_inflight_max;
1755 if ((long)bwnd < tp->t_maxseg * 2)
1756 bwnd = tp->t_maxseg * 2;
1757 tp->snd_bwnd = bwnd;